$A^2$-DP: Annotation-aware Data Pruning for Object Detection

Zengran Wang, Yanan Zhang, Yunlong Qi, Jiaxin Chen, Zehua Fu, Di Huang

Keywords: Object Detection, Data Pruning, Active Learning

Abstract:

As the size of datasets for training deep neural networks expands, data pruning has become an intriguing area of research due to its ability to achieve lossless performance with a reduced overall data volume. However, traditional data pruning usually demands complete dataset annotations, incurring high costs. To tackle this, we propose an innovative Annotation-Aware Data Pruning paradigm tailored for object detection, dubbed
as $A^2$-DP, which aims to reduce the burdens of both annotation and storage. Our approach, consisting of two phases, integrates a hard sample mining module to extract crucial hidden objects, a class balance module to identify important objects in rare or challenging classes and a global similarity removal module that enhances the elimination of redundant information through object-level similarity assessments.
Extensive experiments on 2D and 3D detection tasks validate the effectiveness of the $A^2$-DP, consistently achieving a minimum pruning rate of 20\% across various datasets, showcasing the practical value and efficiency of our methods.

Primary Area: applications to computer vision, audio, language, and other modalities

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 4554